# Challenges in the Classification of Cardiac Arrhythmias and Ischemia Using End-to-End Deep Learning and the Electrocardiogram: A Systematic Review

## TL;DR

This paper reviews challenges in using deep learning with ECGs to detect heart rhythm and blood flow issues, highlighting gaps in current methods and suggesting areas for improvement.

## Contribution

The study systematically identifies 18 new challenges in ECG-based deep learning for cardiac classification, including novel issues in preprocessing, metrics, and explainability.

## Key findings

- Fifty-three challenges were identified in ECG classification using deep learning, spanning preprocessing, models, and evaluation.
- Eighteen new issues were discovered, including limitations in explainability metrics and model confidence assessment.
- Key challenges include complex ECG patterns, comorbidities, and multi-lead data handling.

## Abstract

Background: Cardiac arrhythmias and ischemia are increasingly problematic worldwide because of their frequency, as well as the economic burden they confer. Methods: This research presents a systematic literature review (SLR), based on the PRISMA 2020 statement, that looks into the difficulties in their classification using end-to-end deep learning (DL) techniques and the electrocardiogram (ECG) from 2019 to 2025. A total of 121 relevant studies were identified from Scopus, Web of Science, and IEEE Xplore, and an inventory was created, categorized into six facets that researchers apply in DL studies: preprocessing, DL architectures, databases, evaluation metrics, pathologies, and explainability techniques. Results: Fifty-three challenges were reported, divided between end-to-end DL techniques (15), databases (18), pathologies (9), preprocessing (2), explainability (8), and evaluation metrics (1). Some of the complications identified were the complexity of pathological manifestations in the ECG signal, the large number of classes, the use of multiple leads, comorbidity, and the presence of different factors that change the expected patterns. Crucially, this SLR identified 18 new issues: four related to preprocessing, three related to end-to-end DL, one to databases, one to pathologies, four to metrics, and five to explainability. Particularly notable are the limitations of current metrics for assessing explainability and model decision confidence. Conclusions: This study clarifies all these limitations and provides a structured inventory and discussion of them, which can be useful to researchers, clinicians, and developers in enhancing existing techniques and designing new ECG-based end-to-end DL strategies, leading to more robust, generalizable, and reliable solutions.